synthia-mcp
Official# Synthia
[](https://github.com/synthesizer-project/synthia/actions)
[](https://github.com/synthesizer-project/synthia/blob/main/CONTRIBUTING.md)
[](https://github.com/pre-commit/pre-commit)
[](https://github.com/astral-sh/ruff)
[](https://www.gnu.org/licenses/gpl-3.0)
[](https://pypi.org/project/cosmos-synthia/)
[](https://pypi.org/project/cosmos-synthia/)
<img alt="synthia_logo" src="https://github.com/user-attachments/assets/49c685bd-69d8-4b65-bba1-d78c4813e76d" align="right" width="140px"/>
Synthia is an agent companion for working with [Synthesizer](https://github.com/synthesizer-project/synthesizer).
Synthia bundles an Agent Skill with a local MCP server so Claude Code, OpenCode, and other MCP-capable agents can answer Synthesizer questions from the user's actual installation, and from the project's [Syndex data catalogue](https://synthesizer-project.org/syndex), rather than inflating context and relying on incomplete local memory. Note that it is not designed as a replacement for a human user but simply aims to make learning and using Synthesizer a frictionless experience. For more on these mechanisms see the [MCP documentation](https://modelcontextprotocol.io/docs/2026-07-28/getting-started/intro) and, for Agent Skills, [agentskills.io](https://agentskills.io/home).
## Installation
Synthia requires Synthesizer to be installed first. You'll need to install Synthia in the same environment containing Synthesizer, and you'll need to launch your agent from that same environment.
For agent-assisted installation, send your coding agent this prompt:
```text
Read https://raw.githubusercontent.com/synthesizer-project/synthia/main/INSTALL.md and install Synthia for me.
```
Or install it directly (which is essentially all the install prompt does):
```bash
pip install cosmos-synthia
synthia-install
```
To install an unreleased change, point pip at the repository instead:
```bash
pip install 'git+https://github.com/synthesizer-project/synthia.git'
```
### In case you care what the install does...
`synthia-install` links the bundled skill at `~/.claude/skills/synthia` (which
OpenCode also loads, so one copy serves both clients), registers `synthia-mcp`
with Claude Code when the `claude` CLI is available, and adds it to
OpenCode's `~/.config/opencode/opencode.json`. `--dry-run` shows what it would
change, `--uninstall` reverses it. Restart your client afterwards: skills load
per session, but MCP servers only load at client start.
## Usage
Synthia has no interface of its own. Once installed, ask your agent about
Synthesizer in plain language: the skill triggers on the subject matter, and
the agent calls the MCP tools when it needs an exact local fact, a published
catalogue record, or a support operation.
```bash
cd ~/my-analysis
claude # or opencode
```
```text
> Compare the spectra of a parametric stellar population with little and with
> heavy dust attenuation, and plot them.
```
### What sort of thing can Synthia help with?
- _"Do any of my local grids cover the metallicity range I need, and do they contain H-beta?"_ — answered from the actual grid files you have, not from memory and without delving into the online documentation.
- _"Is there a published BPASS grid with emission lines, and does it reach Z = 1e-5?"_ — answered from the [Syndex catalogue](https://synthesizer-project.org/syndex), including axes, spectra, lines, size, checksum and citations, without downloading the file. Synthia then hands you the `synthesizer-download` command to run yourself.
- _"Why is `spectra['total']` a KeyError after I built a TotalEmission model with fesc=0?"_ — a documented Synthesizer emission model initialisation trap.
- _"Show me what my BPASS v2.2.1 grid's spectra look like at 10 Myr and Z=0.01."_ — renders a figure with known plotting methods and returns its path.
- _"Adapt this script to the Synthesizer version I actually have installed."_ — take outdated scripts and update them for your specific Synthesizer install.
- _"Generate a young and old galaxy with black holes and produce plots showing whether JWST can detect the AGN contribution for a set of reasonable redshifts and galaxy properties."_ — render a complex astrophysics question by applying its knowledge of the local installation and available grids without the need to build up the quick script yourself.
Synthia is at its most useful where an answer depends on your specific installation: which version, which grids, which spectra those grids hold — and on what the project has published that you do not have yet. It is least useful for pure astrophysics questions not specific to Synthesizer, which the model can answer without it.
Synthia can also answer questions about the [Syndex catalogue](https://synthesizer-project.org/syndex), the project's database of published grids, instruments and test data: what exists, what a grid contains, and which release to take. Downloading stays with `synthesizer-download`, which Synthia gives you the command for rather than running.
## Performance
The benchmarks directory contains a series of tests for common Synthia prompts. These were run on a bare agent (i.e. no plugins, MCPs, or skills) using Sonnet 5 with and without Synthia. Below is a comparison of the usage and runtime, measured over the 15 cases that predate the catalogue tools.
| metric | baseline | synthia | change |
| ----------------- | -------- | ------- | ------ |
| total cost | $11.30 | $8.22 | −27% |
| total wall time | 3682 s | 2174 s | −41% |
| source read | 954 kB | 92 kB | −90% |
| cache-read tokens | 21.5 M | 14.9 M | −31% |
| turns | 443 | 341 | −23% |
Note that every Synthia script produced a correct result or a runnable script while baseline did in all but one case. Synthia is designed to be correct more often but importantly it'll get to the correct answer with less resource and often faster.
## Development
See [CONTRIBUTING.md](https://github.com/synthesizer-project/synthia/blob/main/CONTRIBUTING.md) for the full development workflow.
## Licence
[GNU General Public License v3.0](https://github.com/synthesizer-project/synthia/blob/main/LICENSE).
TDQS
Scored across 13 tools
Most tools are clearly distinct: search vs list vs inspect vs plot families each target different resources and actions. Minor overlap exists between search_documentation and find_example for how-to questions, and the three plot_grid_* tools share a prefix but are differentiated by what they render.
Tool names follow a mostly consistent verb_noun snake_case pattern (search_, list_, inspect_, plot_, describe_, validate_). The main inconsistency is using 'find' vs 'search' for similar lookup actions and 'describe' vs 'inspect' for catalogue vs local metadata, but the pattern remains readable and predictable.
13 tools is well-scoped for a server covering documentation search, examples, catalogue browsing, local grid inspection/plotting, environment/API inspection, and script validation. Each tool has a clear role and none feel redundant or like filler.
The tool surface covers the main workflows: finding docs/examples, discovering and describing catalogue datasets, listing and inspecting local grids, plotting grid outputs, inspecting the environment/API, and validating scripts. Minor gaps exist, such as no tool to return raw spectral/line data as structured values rather than plots, but these are workable limitations rather than dead ends.